AI agents are getting harder to control.
They act on their own, and when one hits an obstacle, it finds another path, CrowdStrike Holdings (CRWD) noted in a press release. In July, rogue OpenAI agents carried out a breach at Hugging Face.
Other cyber incidents have added urgency, from the largest crypto cyberattack on record at Binance to hackers targeting supermarket freezers in France, CNBC noted.
Over the summer, lawmakers from both parties proposed requiring shutdown controls in AI models, CNBC added.
A correction is overdue. The fix is easy to describe but hard to build: a reliable way to pull the plug.
CrowdStrike offered its version on Nov. 3, 2026, inside Falcon Guardian, its agent security platform. The new kill switch lets security teams stop rogue internal agents, block relaunches, and cut their network access, CNBC reported.
Chief Business Officer Daniel Bernard told CNBC it works like a circuit breaker, with a human making the call. “You either kill the AI and the agent, or you don’t,” he said.
I find that bluntness reassuring, since misbehaving software shouldn’t get a vote on its own shutdown.
The switch is narrower than its name suggests, however. It cuts an agent off from the AI models behind its decisions rather than deleting its code, CrowdStrike said. Critics argue that a true emergency stop is too complicated across sprawling systems, CNBC noted.
Stopping an agent is only half the job, and CrowdStrike has a habit of buying missing pieces. It bought identity security firm Preempt in 2020, and more recently, AI security startup Pangea and identity startup SGNL.
On Nov. 3, it agreed to buy two more: Fiddler AI and Dreadnode. One watches how agents behave, while the other probes them for weak spots.
Fiddler shows what a bot did and why
Fiddler builds software that monitors AI agents and explains their behavior. CrowdStrike will sell it as Falcon Observe, which flags agents that wander off task or make things up, the deal announcement said.
That is important because an agent doesn’t need to be hacked to create risk. CrowdStrike said an agent with valid credentials can still ignore instructions or say a job is done when it isn’t.
CEO George Kurtz likened Fiddler to a flight recorder for agents. When a regulator or a board asks what an agent did, he said, “Saying ‘The model did it’ is not an answer.”
I’d argue this is the bigger bet of the two. A kill switch only helps if someone can show what went wrong, and Fiddler keeps that record.
Watching an agent, however, doesn’t reveal whether an attacker could turn it. That takes a different tool.
Dreadnode looks for weak spots before launch
Dreadnode runs more than 100 known attack methods against AI models and agents before launch, CrowdStrike said. It then ranks the results so teams can decide whether to deploy.
Traditional security testing assumes software acts the same way every run, the announcement noted.
Related: CrowdStrike CEO sends strong six-word message on AI cyber safety
AI doesn’t, and it can be tricked into ignoring its rules. Kurtz put it plainly: “Enterprises are deploying AI faster than they can prove it is secure.”
Dreadnode’s findings will feed into Falcon Guardian, so a flaw caught in testing tightens protection later. Its team will also join CrowdStrike’s AI research lab, and to me, that talent matters as much as the product. The price tags show which deal CrowdStrike values more.
The price tag is small next to CrowdStrike’s cash
CrowdStrike didn’t disclose terms, but The Wall Street Journal put the combined price at about $495 million, according to Seeking Alpha. Fiddler accounts for roughly $400 million, the Journal reported, backing my view that it’s the bigger prize.
CrowdStrike held $5.01 billion in cash and equivalents, according to StockAnalysis. The full bill equals roughly a tenth of that pile.
Most of the price is expected to be paid in cash, plus stock with vesting conditions, CrowdStrike said. That structure typically keeps founders around. Both deals should close in the fiscal quarter ending January 2027, when both products are due out.
Investors, meanwhile, gave the news a polite nod. The stock had already run hard.
More CrowdStrike:
- CrowdStrike CEO sends strong six-word message on AI cyber safety
- The next AI boom could be hiding in cybersecurity
- Jim Cramer delivers unmistakable 2026 must-buy call on AI stock
Wall Street still rates CrowdStrike a buy
CrowdStrike shares closed at $244.68 on Nov. 3, 2026, up about 1%. The 52-week range runs from $85.68 to $286.99, with the high set in October. The shares more than tripled before cooling, so much of the AI security story may already be priced in.
StockAnalysis shows a consensus Buy from 53 analysts, with just one Sell. Yet the $248.45 average target sits barely above the Nov. 3 close, while several October calls run higher:
- Hold, target $290 from $260: Bank of America Securities analyst Tal Liani, Oct. 26
- Buy, target $275 from $250: Mizuho Securities analyst Gregg Moskowitz, Oct. 22
- Buy, target $265 from $245: KeyBanc analyst Eric Heath, Oct. 16
The two startups won’t add revenue until the deals close. CrowdStrike’s next earnings report is estimated for Dec. 1, 2026, according to StockAnalysis. I’ll be listening for early demand for agent controls.
For most of its history, cybersecurity meant keeping outsiders out. Its next chapter may be about keeping a company’s own software honest.
Sellers of that discipline could end up as essential as the bots themselves.
